Guides

When two answers look right

Updated 20 Nov 2026. Unofficial guide, checked against the official exam guides.

The hardest items on the Claude exams are not the ones about a feature you’ve never heard of. They are the ones where two options would both work. Szymon Paluch, who passed Architect Foundations in July 2026, describes it exactly: the questions that took him longest had several reasonable answers, and a detail in the scenario decided.

That is by design. The guides describe the items as scenario-based, and the new Associate guide says they assess judgment in work scenarios. A good wrong option sounds like something a competent person might do. Your job is to find the sentence in the scenario that rules them out.

Where the deciding constraint hides

When you’re stuck between two options, go back to the scenario and look for these, in this order.

Words that make something mandatory. Must, always, never, every, required, auditable. An instruction in a prompt makes behavior likely, while a hook, a permission rule, a schema or a code path makes it certain. When the scenario says a step must happen every time, the option that relies on prompt wording is the trap, however well it’s worded.

The number. A latency target, a cost ceiling, a volume, a deadline. An option can be good practice and still break the number. Take a check that has to finish before a developer can merge. Sending it through a batch API is a sound cost saving for overnight work and wrong here, because a batch can take hours. If the scenario gives you a number, at least one option breaks it.

What hasn’t changed. “The instructions have not been edited.” “The tool definitions are the same as last month.” These lines are there to eliminate options. If output got worse and the instructions are unchanged, rewriting the instructions can’t be the fix. One of the published Associate sample questions works exactly this way. A Project’s drafts degrade while its instructions and Skill stay unchanged, which leaves the stale knowledge file and the accumulated memory as the only layers that can explain the change.

The evidence. Logs, traces, metrics, a latency breakdown. If the scenario shows that output generation takes most of the time, an option that speeds up the start of the response is solving a problem the evidence doesn’t show. Look for the option that matches the measurement, not the one that improves things in general.

Who will see it and what an error costs. An internal draft and a client deliverable call for different levels of review. An action that can be reversed and one that can’t call for different gates. The stakes tell you how much checking is enough and where a human must stay in the loop.

The upgrade trap

One kind of wrong option turns up in practice questions for every exam: the general upgrade. The course cheat sheets for all five exams list a version of it. Use a bigger model, a longer context window, more retries, a stronger warning in the system prompt, a longer cache lifetime. Each sounds safe, and each is wrong when the scenario has already shown you a specific cause.

If the cache hit rate is near zero, a longer cache lifetime does nothing while something keeps changing the prompt prefix, such as a timestamp near the top. If a value is missing from the source document, retrying the extraction won’t produce it, and the right move is to return nothing and route the case for review. The specific fix that addresses the evidence beats the general one that might help.

Questions that ask what to do first

“Which step should the team take first?” is a question about order, so every option may be something you’d eventually do. Three rules help:

  1. If harm is happening now, stop it first. A customer who asks for a human gets one now, before you investigate their case. A tool that is leaking data gets locked down before anyone writes the incident report.
  2. Otherwise, establish the cause before you change anything. A problem that started right after a release looks like a reason to roll back. If the scenario hints that only part of the traffic is affected, first check which version that traffic was on. A fix made before you know the cause can remove the evidence along with the symptom.
  3. Then take the cheapest step that targets that cause. If tools are being chosen wrongly, improve their descriptions before you build a routing classifier. If one field is extracted badly, look at that field before you reconsider the whole pipeline.

Paluch gives two shapes this takes. One asks for the first step in a given situation. The other offers three technically correct solutions and asks which suits the cost and latency requirements best. Both are answered by the scenario, not by which option is the best practice in general.

Practise the elimination, not just the answer

When you review a practice question, don’t stop at the correct option. For each wrong option, find the sentence in the scenario that rules it out. If you can’t find one, you got lucky, and the same pair will cost you on the exam. That is why every explanation in the free forms says why each wrong option is wrong, not only why the right one is right.

All guides Changes log